OpenAI has announced new agentic capabilities for ChatGPT, which claim to bring together three strengths of earlier breakthroughs: the Operator’s ability to interact with websites, deep research’s skill in synthesising information, and its intelligence and conversational fluency.
Agentic capabilities
With ChatGPT’s new agentic capabilities, users can instruct it to perform tasks and manage complex workflows. To keep users in control, ChatGPT requests permission before taking actions of consequence. Users can also easily interrupt, take over the browser, or stop tasks at any point.
The ChatGPT agent can access connectors, enabling it to integrate with workflows and access relevant, actionable information. Authenticated connectors allow ChatGPT to see information and execute tasks. Users can also schedule completed tasks to recur automatically.
Safeguarding security
With new risks that come with ChatGPT taking actions on the web, OpenAI have strengthened the controls from Operator’s research preview and added safeguards for challenges such as handling sensitive information on the live web, broader user reach, and (limited) terminal network access.
The company has also trained and tested the agent on identifying and resisting prompt injections, and has implemented mitigations to address model mistakes. OpenAI announces the launch is just the beginning as the company continues to add significant improvements to ChatGPT.
